Computers and Chemical Engineering 81 (2015) 200–217

Contents lists available at ScienceDirect

Computers and Chemical Engineering

journal homepage: www.elsevier.com/locate/compchemeng

Conceptual design and optimization of chemical processes under

uncertainty by two-stage programming
J. Steimel ∗ , S. Engell
Process Dynamics and Operations Group, Department of Biochemical and Chemical Engineering, TU Dortmund, Emil-Figge-Str. 70, 44227 Dortmund,
Germany

a r t i c l e i n f o a b s t r a c t

Article history: This contribution presents a method and a tool for modelling and optimizing process superstructures in
Received 4 November 2014 the early phase of process design where the models of the processing units and other data are inaccurate.
Received in revised form 10 May 2015 To adequately deal with this uncertainty, we employ a two-stage formulation where the operational
Accepted 15 May 2015
parameters can be adapted to the realization of the uncertainty while the design parameters are the
Available online 27 May 2015
first-stage decisions. The uncertainty is represented by a set of discrete scenarios and the optimization
problem is solved by stage decomposition. The approach is implemented in the computer tool FSOpt (Flow
Keywords:
sheet Superstructure Optimization) FSOpt provides a flexible environment for the modelling of the unit
Chemical processes
Process design
operations and the generation of superstructures and algorithms for the translation of the superstructure
Mathematical modelling into non-linear programming models.
Optimization The approach is applied to two case studies, the hydroformylation of dodec-1-ene and the separation
Uncertainty of an azeotropic mixture of water and formic acid.
Evolutionary algorithm © 2015 Elsevier Ltd. All rights reserved.

1. Introduction the point in time when the information is sufficient is usually dif-
ficult. In this work, it is assumed that no further experiments have
In this contribution, process synthesis is understood as the to be scheduled when the uncertainty in the parameters (obtained
choice of the best options from a set of promising design candidates from the available experimental data) has no influence on the rank-
during the early stage of process development. This decision has ing of the design alternatives any more so that options that are not
to be made under incomplete information, as not all alternatives promising do not have to be pursued further and the effort can be
have been explored in experimental work on the laboratory and focused on one or few design alternatives. In design optimization,
pilot plant level and the models of some operations are inaccurate, both design parameters and operational parameters must be spec-
also information on cost and prices is usually tentative. The need ified to evaluate a design. However, while the design parameters
for explicitly including uncertainty into the decision making pro- cannot be changed any more after the plant has been built and the
cess for process synthesis has already been discussed by Grossmann true behaviour of the different units is observed.
and Sargent (1978) and Acevedo and Pistikopoulos (1996). The operating parameters can be adapted to the realization of
The amount of information that is available during the early the uncertainties. Obviously, this potential for recourse should be
process synthesis phase is usually rather limited and the process taken into account in the selection process which is why we employ
models are of different levels of quality. For some process steps a two-stage optimization approach where the design parameters
only semi-quantitative information may be available. The aim of e.g. the number of trays of a column or the size of a reactor are
our work is to find the best structures in the presence of these the first stage variables while the operating parameters, e.g. the
uncertainties using mathematical programming techniques. reaction temperature or the reflux ratio of a column are the recourse
During the development of a new process, one of the main chal- parameters that are adapted to the uncertainties. The uncertainty is
lenges is to stop the laboratory or pilot plant experiments as soon represented in the optimization by a finite set of discrete scenarios.
as sufficient information has been gathered. The quantification of The approach is implemented in the computer tool FSOpt (Flow-
sheet Superstructure Optimization) using a stage decomposition
approach to solve the resulting optimization problems. The first
∗ Corresponding author. Tel.: +49 231 755 5341; fax: +49 231 755 5129. stage variables (design decisions and design parameters) are opti-
E-mail address: jochen.steimel@bci.tu-dortmund.de (J. Steimel). mized by an evolutionary algorithm. The design candidates are

http://dx.doi.org/10.1016/j.compchemeng.2015.05.016
0098-1354/© 2015 Elsevier Ltd. All rights reserved.
 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217 201

evaluated after optimizing the operational parameters using a NLP Several methods for expressing superstructures as mathemati-
solver. This is similar to the solution technique for mixed-integer cal programming problems exist. Classically, superstructures are
two-stage stochastic programmes proposed in Till et al. (2007) and modelled as mixed-integer non-linear programmes (MINLP) in
Tometzki and Engell (2010). which the integer decisions are relaxed using Big-M formulations
The models and the superstructures are described using the (Grossmann et al., 2000). A newer approach for modelling super-
language Modelica, and executable code for optimization is auto- structures is Generalized Disjunctive Programming which is based
matically generated from the object-oriented models within the on logic propositions and disjunctions (Yeomans and Grossmann,
tool FSOpt. This provides the user with an abstract way to formulate 1999; Lee et al., 2003). This approach allows the grouping of equa-
the models without having to deal with the details of the implemen- tions into disjunctive sets, for which efficient relaxations can be
tation in the actual solver. The availability of high-level constructs calculated using a convex-hull approximation. With better relax-
like objects, inheritance and connections increases the chance for ations the optimization performs faster, as larger parts of the
model reuse and easier maintenance due to better readability and tree can be pruned during the search. A problem in MINLP-based
separation of concerns. approaches is the need to solve the relaxed root problem. If this
The modelling concept of FSOpt and the model transforma- problem is large and difficult to solve, most MINLP algorithms will
tions by symbolic algorithms are described and we demonstrate fail to solve the overall optimization problem as no starting point
the capabilities of the tool FSOpt in its present extended form for can be found.
two challenging case studies. A recent approach for modelling block-oriented superstructures
The first case study features the design of a process for the hydro- (Friedman et al., 2013), exploits the hierarchical, highly structured
formylation of dodec-1-ene, which has already been discussed in topology that is inherent in most superstructures. This approach
previous work (Steimel et al., 2013, 2014). In this contribution the has been implemented in the Pyomo modelling framework (Hart
model formulation has been improved and new uncertainties are et al., 2011).
considered. One difficulty of solving mathematical programming problems
The second case study deals with the separation of a mixture that arise from superstructure optimization is due to the integrality
of formic acid, water, methyl-formate and methanol which is part of some decision variables, leading to huge disjoint search spaces.
of the Kemira–Leonard process for the production of formic acid. The approach in Dowling and Biegler (2015) aims at finding better
In this case study the presence of an azeotrope between water relaxations than those that result from the use of standard MINLP
and formic acid necessitates the use of non-ideal thermodynamics techniques. For distillation columns the relaxation is based on a
which are implemented using the Wilson equation. The complex physically realizable tray bypass.
thermodynamic equilibrium calculations pose challenges for the A second major problem is the existence of (often many) local
modelling and the optimization as discussed below. minima in flowsheet optimization problems (Urselmann et al.,
The article is structured as follows. The theoretical background 2011). This calls for the use of rigorous global optimization algo-
and the state of the art concerning the optimization of super- rithms which however are not yet able to solve problems beyond
structures and modelling the uncertainties are summarized briefly the optimization of a single apparatus (Ballerstein et al., 2011).
in Sections 2 and 3. Section 4 provides a short overview of two- When uncertainties are considered by modelling them as a num-
stage stochastic programmes and related solution techniques. In ber of discrete scenarios, the problem size increases proportional
Section 5 the main features of the tool FSOpt are discussed. Sec- to the number of scenarios. This further increases the difficulty
tion 6 describes the first case study and highlights some interesting of solving the resulting optimization problems. In the context of
results. The second case study is presented in Section 7. The article the two-stage stochastic optimization formulation, we employ a
closes with Section 8, which consists of a discussion, some conclud- stage decomposition approach so that the scenario sub-problems
ing remarks, and an outlook on future research. can be solved by standard NLP solvers. The first stage problems are
highly non-convex with disjunct feasibility sets and we use an evo-
lutionary algorithm to generate good, though not provable optimal
2. Superstructure optimization using mathematical solutions.
programming

An overview of mathematical programming for process synthe- 3. Handling uncertainty in design optimization
sis is given in Grossmann et al. (1999). Despite remarkable progress
in recent years, there are still many open problems. One of them, The seminal article by Grossmann and Sargent (1978) discussed
related especially to the early design phase, is that algorithm-based the need to include uncertainty into the decision making process
short cut models are not suitable for rigorous optimization. in process optimization. Since then, many approaches have been
Secondly, the size of the flowsheets and the level of detail of the presented in the literature, four of which will be described in the
models are still limited. Examples reported in the literature cover following section.
flowsheets exhibiting one or two distillation columns (c.f. Dowling The goal of parametric programming (Acevedo and
et al., 2014; Dowling and Biegler, 2015), auxiliary separation units Pistikopoulos, 1996) is to find a description of the solutions
(Skiborowski et al., 2014) or reaction steps (Recker et al., 2015). If of the design problem depending on the uncertainties in a closed
the level of detail of the model increases, robust convergence of the form. The idea to map the uncertainty space to the solution
solution is still difficult to achieve and requires sophisticated semi- space using single function is intriguing, but so far limited to few
automated initialization techniques (Dowling and Biegler, 2015; uncertain parameters, due to the need of repeated optimizations
Skiborowski et al., 2014). If complex networks are synthesized the of a parametric master problem and NLP sub-problems. Another
underlying models are often simplified (Quaglia et al., 2014) and/or approach which does not require explicit sampling of the uncer-
only models of a single kind of unit are considered. Examples for tainty space was presented by Pistikopoulos and Ierapetritou
this can be found in water treatment network synthesis (Tanvir and (1995). For uncertainties that are described by continuous proba-
Mujtaba, 2008; Khor et al., 2012). bility distribution functions a two-stage stochastic programming
In most of these examples the design problem is modelled as problem is formulated which considers economic aspects as well
a superstructure, which is a graph containing all combinations of as process feasibility. Another example of process optimization
process units and connections that are considered. without scenario sampling is given in Rooney et al. (2003), in which
202 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217

the feasible region under uncertainty is approximated by a four- is based on the fact that in early process design many parameters
level optimization problem, yielding a single, smooth non-linear are simply not known precisely. The uncertainties are represented
programme representing the process feasibility problem. by a set of scenarios with fixed values of the uncertain parameters
In contrast to the parametric approaches, sampling-based either at the upper or lower limits or at the nominal values. The
approaches use a representative number of samples (called sce- probability of these values is assumed as equal in the case studies,
narios or realizations) that are taken from the uncertainty space to and a full-factorial sampling is used to compute the average value
approximate the distribution of the cost function as a function of of the cost function.
the uncertain variables and to optimize this distribution in some Three main approaches exist for solving two-stage stochastic
sense (usually the expected value is minimized or maximized). The programmes. The straightforward approach is to solve the deter-
number of samples necessary for a sufficient representation of the ministic equivalent of the stochastic problem as a large monolithic
probability distribution of the objective function strongly increases mixed-integer linear or non-linear problem (MINLP). This approach
with the number of uncertain parameters. is explained in Birge and Louveaux (1997). The monolithic formu-
Several sampling techniques have been discussed by Diwekar lation leads to weakly connected blocks in the incidence matrix in
and Kalagnanam (1997) and Helton and Davis (2003) with the goal the problem as the model equations are copied for each scenario,
of reducing sample size while still achieving high fidelity in rep- leading to a very sparse problem. The sparsity can be exploited by
resenting the probability distribution. In process design, often the efficient solvers. Obviously the number of variables and of equa-
uncertainties cannot be quantified by distributions but only inter- tions increases roughly proportional to the number of scenarios
vals for the expected values are available. If only the bounds of the that are considered.
intervals and the nominal value are taken into account, one obtains The structure of the problem due to the existence of weakly
information about the best and worst cases profitability which can coupled blocks can be exploited using decomposition approaches.
be sufficient for the selection of one or a small number of design Caroe and Schultz (1998) proposed to use scenario decomposition
alternatives. to solve mixed-integer linear two-stage stochastic programming
problems. Here, the original problem is decomposed along the sce-
4. Two-stage stochastic programming narios by introducing copies of the first stage variables in each
scenario so that the scenarios can be solved independently and
Two-stage stochastic programming is one of the main using a Lagrangian relaxation of the condition that the variables
approaches for optimization under uncertainty (Sahinidis, 2004). that shared by the scenario solutions, i.e. the first stage variables
As discussed above, process synthesis and design problems with must be equal. The drawback of this approach is the difficulty of
uncertain parameters naturally lead to two-stage stochastic prob- solving the Lagrangian dual problem and of searching for the best
lems as the design parameters have to be fixed while the operating feasible solution for the integer first stage variables.
parameters can be adapted to the realization of the uncertain The third approach, e.g. used in Till et al. (2007) proposes a stage
parameters. The core idea behind two-stage stochastic program- decomposition. A master algorithm performs a search on the first
ming is to optimize the first stage decisions (here the design stage variables and the decoupled sub-problems are solved inde-
parameters) for a suitable cost function that is computed for an pendently, e.g. using standard MILP or NLP solvers. Thus the size
optimal choice of the second stage variables for each scenario. of the sub-problems is much smaller than that of the overall prob-
Thus the option to counteract the uncertainties is anticipated in lem. The difficulty of this approach is that the master problem in
the decision, making the result much less conservative than e.g. a which the solution of the scenario problems is embedded is non-
worst-case design. The design must be feasible for all realizations of convex and may have non-connected feasibility regions. Therefore
the uncertainty, however only after adaptation of the second stage stochastic solution techniques are most appropriate for its solution.
variables (operational parameters).
The general form of a two-stage-stochastic optimization prob-
5. The tool FSOpt
lem is given in Eq. (1)–(5).


˝ The two-stage-stochastic approach to design optimization was
min G(y) + ω Fω (y, xω ) (1) implemented in the tool FSOpt, first described in Steimel et al.
y,xω
ω=1 (2013) and Steimel et al. (2014). FSOpt provides an interface for
modelling the process steps using the Modelica language (Olsson
s.t. g(y, xω ) ≤ 0 (2)
and Otter, 2012). The object-oriented language Modelica has a
f (y, xω ) = 0 (3) well-defined, open grammar, enabling the implementation in non-
commercial academic software and the ability to define declarative
xL,ω ≤ xω ≤ xU,ω (4)
model descriptions using a high-level syntax, improving model
yL ≤ y ≤ yU (5) reusability. Modelica is mostly used as a modelling language in the
automotive industry and the energy sector, but there is also a grow-
The objective function (Eq. (1)) is composed of two parts. The
ing interest in process systems engineering (c.f. Holmgren, 2003;
first term G(y) describes the cost incurred by taking the first-stage
Baharev and Neumaier, 2012).
decisions y, which can be non-linear in the general case. The second
FSOpt automatically translates the flowsheet superstructure
term is the expected value of the second-stage costs, approximated
model into a two-stage-stochastic programming model. The flexi-
via a summation of the costs for ˝ discrete scenarios (Fω ) multi-
ble transformation tool chains in FSOpt enables the connection to
plied by the probabilities of the scenarios ω . The design problem is
several solution approaches and back-end solvers.
subject to inequality constraints (Eq. (2)). In design problems, these
commonly arise from product property constraints and capacity
specifications. The problem is constrained further by equality con- 5.1. Model transformation
straints (Eq. (3)) that result from the models of the process units and
the interconnections between them. The design decisions y and the FSOpt employs a symbolic object model to describe hierarchical
recourse decisions xω are subject to upper and lower bounds. optimization problems on which all model transformations oper-
In this contribution, the uncertainties are assumed to be dis- ate. The object model is implemented as a class library in C# and
tributed between an upper and a lower bound. This assumption includes classes for the most common elements encountered in
 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217 203

Fig. 1. Example of the model transformation from RTN to Modelica.

non-linear-optimization problems, like objective functions, vari- by adding additional equations and variables, which cannot be
ables, equations, inequality constraints and bounds. FSOpt includes expressed in the RTN formalism. Alternatively, the modelling
classes for describing algorithmic structures, which means that process can start on the equation level, e.g. when a graphical repre-
custom user functions (e.g. thermodynamic property calculations sentation of the problem would be too complex. For optimization,
commonly required in process simulation) are directly supported
in the language without the need for external functions. Mecha-
nisms for modelling hierarchically composed models are included
supporting the object-oriented concepts of classes, instances and
inheritance. The framework is largely independent of the front-end
language in which the models are formulated. Parsers for subsets of
the Modelica and AMPL languages are implemented in the current
version. After parsing, the model is translated by a process com-
monly referred to as flattening or instancing (for an explanation of
the flattening process with regard to Modelica models, see Mauss
(2005)). The entire transformation work flow as it is implemented
in FSOpt is shown in Fig. 2. The modelling process usually starts on
the flowsheet level, on which the process units and connections are
modelled using the Resource-Task-Network (RTN) (Kondili et al.,
1993) formalism. In FSOpt, a graphical editor for a variant of RTN
is implemented in which tasks are used to model unit operations,
resources represent the models which can be used to implement a
given unit operation, states are equivalent to material streams and
arcs represent stream connections between units. An example of
this transformation is given in Fig. 1.
In this figure, the black circles denote streams of material
(states), the black-bordered white boxes denote unit operations
(tasks) and the grey rounded boxes represent resources (models or
requirements). Each resource can be enriched by several attributes
(in this case model parameters). In the example shown in Fig. 1,
the unit operation Reaction can be implemented by two exclusive
alternative models, either as a plug-flow-reactor (with model RPFR)
or as a continuously-stirred-tank-reactor (with model RCSTR). This
exclusive choice is modelled by a binary variable Y that is used for
the conditional instantiation of the respective reactor sub-model
in Modelica.
An automatic reformulation then transforms the RTN into a
Modelica model, which can then be further modified manually, Fig. 2. The model transformation work flow in FSOpt.
204 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217

Table 1 algorithm does not provide support for imposing constraints. This
Specification of the hybrid evolutionary algorithm.
capability has to be added by problem-specific operators e.g. either
Element Comment by a penalty-function approach or a repair operator that gener-
Representation Vector of real and discrete values ates feasible values from infeasible proposals. We here formulated
Recombination Discrete random one-point crossover the models in such a way that no constraints on the design deci-
Mutation Gaussian perturbation sions (except of box constraints) are needed and all combinations
Parent selection Uniform random of design parameters yield a first-stage feasible design. All other
Survivor selection (, , )
constraints are handled in the underlying continuous optimization
Specialty Self-adaptation of mutation strength
Local optimization of recourse problem. An exception is discussed in the first case study, where a
custom repair operator is used for the feed location of the distilla-
tion columns.

Algorithm 1. The hybrid evolutionary algorithm

the model is first parsed and then transformed into an instance of
the optimization object model. A number of transformations can be 1 generation=0
applied to this model, namely flattening (recursively resolving class 2 pop = initialize()
instances), instancing (expanding sets and arrays), model reduction 3 localSearch (pop)
(preprocessing, detecting trivial equations, variable aliases, etc.) 4 while terminate = = false do
and finally symbolic differentiation. In the current implementation, 5 parents = select (pop, numOfParents)
symbolic differentiation is not applied to user-defined function 6 children = recombine (parents)
calls, for which a finite difference approximation is used to compute 7 children = mutate (children)
the partial derivatives of the functions. Every step is implemented 8 localSearch (children)
as an independent and loosely coupled component, facilitating easy 9 survivors = select (pop + children)
reconfiguration of the entire work flow. For example, calculating 10 terminate = checkTermination (generation)
first and second derivatives is not necessary when targeting the 11 generation = generation+1
GAMS platform as the computational back-end, so this step can be 12 end
omitted, and the code can be directly generated after the reduction
of the system of equations, reducing the effort for transformation The evolutionary search is enhanced by a local search which
and preprocessing. improves the fitness of the individuals using a gradient-based
solver. The fitness is equivalent to the objective function value of
5.2. Code generation the NLP problem (or the set of NLP problems for all scenarios con-
sidered) that is defined by the design parameters that are encoded
The result of the transformation process is an optimization prob- in the genome of the individual. The algorithm which translates
lem where all object references and set/index instances have been the genome (the vector of design decisions) to a phenotype (a non-
resolved. Generating code from this flat model is straightforward. linear programming problem) is given in Algorithm 2. The functions
Each element in the symbolic object model can be transformed with parse, instantiate, residualize, calcJacobianAndHessian and render are
a mapping into a string of output characters in the desired language. symbolic transformation steps implemented in FSOpt.
The translation algorithm traverses the entire object model until all
elements have been handled. Algorithm 2. The local search algorithm
In FSOpt, the code-generation is implemented using the
StringTemplate software. In the current version FSOpt can gen- Data: a piece of FSOpt-Modelica source code encoding the super-
erate optimization code for GAMS, IPOPT through the C interface, structure
and Matlab. This makes FSOpt tool-independent and formalism- Result: a ProblemResult object containing the numerical results
agnostic. Using FSOpt, it is possible to develop an abstract model 1 optimProb = parse (sourceCode)
and to use the model in different implementations in a number of 2 optimProb = instantiate (optimProb)
other tools, choosing the one which is best suited for the particular 3 optimProb = residualize (optimProb
task. 4 optimProb = calcJacobianAndHessian (optimProb)
5 executableCode = render (optimProb)
5.3. The optimization algorithm 6 result = execute (executableCode)
7 return result
The evolutionary algorithm that is used to explore the space
of design decisions is a standard algorithm that was developed in The parse function translates the Modelica model of the
our group based on experience described in Till et al. (2005), Till superstructure into an internal data structure that describes a hier-
et al. (2006) and Urselmann et al. (2011). Its classification according archical NLP. The instantiate function transforms the hierarchical
to the guidelines in Eiben and Smith (2008) is given in Table 1. MINLP into a flat NLP that consists only of equations and vari-
The survivor selection is realized using a (, , )-selection, which ables and removes all high-level modelling constructs like loops,
means that the population size is set to  and in each generation  conditionals and connections. The residualize function transforms
new individuals are created. From the  +  individuals the  best all equations into residual form, then the calcJacobianAndHessian
are selected, provided that they are not older than  generations. function symbolically derives all partial derivatives for the model
All evaluated solutions are stored and are available for inspection equations regarding all variables. The render function transforms
after the algorithm has terminated. the flattened data structure into executable code in the desired
The evolutionary cycle, as encoded in FSOpt, is shown in target language. The executable code contains an execute function,
Algorithm 1. It is a repeating sequence of selection, recombination, which performs the two-stage stochastic optimization with stage
mutation and again selection. Termination criteria are a maximum decomposition. The pseudo-code of the execute function is given in
run-time, maximum number of generations, and maximum num- Algorithm 3, but the actual implementation varies depending on
ber of stalled generations. The standard version of the evolutionary the target language.
 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217 205

Fig. 3. Flow sheet superstructure.

Algorithm 3. The execute function temperatures and a low miscibility at lower temperatures. The
reaction is typically performed in the homogeneous phase at higher
Result: a ProblemResult object containing the numerical results temperatures, while the phase separation takes place at lower
1 result = new ProblemResult() temperatures. The process involves several chemical components,
2 for i = 1 to number of scenarios do dodec-1-ene (nC12), dodec-2-ene (iC12), n-tridecanal (nC13),
3 setScenario (i) 2-methyl-dodecanal (iC13), decane (C10), di-methyl-formamide
4 status = solveProblem() (DMF) and the catalyst. The thermomorphic solvent system is
5 result = updateResults (result, status) formed by C12 and the solvents C10 and DMF.
6 end
7 return result
6.2. Model description

6. Case study 1: hydroformylation of dodec-1-ene The reactor is modelled as a CSTR based on the kinetic model
in Hentschel et al. (2014). The distillation columns are modelled
6.1. Process description assuming equilibrium stages and an ideal behaviour of the liquid
and the vapour phase. The feed location and the total number of
The case study analyzed in this section is taken from the trays in a column are design parameters. The standard hybrid evo-
ongoing work of the DFG Collaborative Research Center SFB/TR lutionary algorithm presented before was extended by a special
63 InPROMPT. In the past years, InPROMPT has focussed on the safeguard in the first-stage optimizer which was used to guaran-
development of process concepts for the hydroformylation of tee that the feed tray location is always smaller than the maximum
long-chain alkenes using innovative solvent systems, in particular number of trays of a column. This was implemented as a local repair
the hydroformylation of dodec-1-ene. The reaction combines algorithm in the mutation function.
synthesis gas and dodec-1-ene and is catalyzed by a homoge- The superstructure of this process is depicted in Fig. 3. The reac-
neous rhodium-based catalyst. The challenge of retaining this tion product is separated into two phases in a decanter. The catalyst
very expensive catalyst can be solved using a thermomorphic is retained in the polar, DMF-rich phase, which is recycled. The apo-
solvent system. Thermomorphic solvent systems, as proposed lar product phase is fed to a solvent column, where the unreacted
e.g. by Behr et al. (2006) employ a mixture of components of feedstock (nC12, iC12) and the solvent (C10) are drawn over top
different polarities which exhibit a high miscibility at higher and recycled to the reactor as well.

Table 2
Decisions of the optimization problem for case study 1.

Process step Variable Type Unit Min Max

Solvent column NStages Design-discrete [–] 15 30

Product column NStages Design-discrete [–] 15 30
Solvent column Feed stage Design-discrete [–] 1 NStages
Product column Feed stage Design-discrete [–] 1 NStages
Crystallizer 1 Exists Design-Boolean [–] 0 1
Crystallizer 2 Exists Design-Boolean [–] 0 1
Reactor Volume Design-continuous m3 5 30
HX Area Design-continuous m2 5 20
Cooling water Price Uncertain kD /kW h 1 . 08−5 1 . 32−5
Steam Price Uncertain kD /kW h 6 . 003−5 7 . 337−5
Decanter K1Cat Uncertain [–] 4 . 5961−5 5 . 6175−5
Reactor T Recourse K 353 383
Reactor pCO Recourse bar 5 20
Reactor pH2 Recourse bar 5 20
Feed nC12 Recourse kmol/h 1 20
Decanter T Recourse K 288 298
206 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217

Fig. 4. Evolution of the fitness function over the generations for the case study 1.

The bottom composition of the solvent column is mostly made of nC13 is required to be at least 99mol− %, and the capacity was
up by the two isomers of tridecanal and a small fraction of dodec- fixed at 10,000 t/y.
1-ene. The purification step consists of a distillation column and The model of the hydroformylation process is formulated using
up to two optional crystallizers. The purification step yields high- discrete decisions, conditional equations and algorithmic function
purity nC13, and a side stream of iC13 mixed with nC13. The purity calls and is solved using the custom hybrid evolutionary algorithm

Fig. 5. Scenario production cost for case study 1.

 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217 207

Fig. 6. Flowsheet for design 6.

described in Section 5.3 which optimizes the design parameters. feed nC12 and the temperature in the decanter. The number of equa-
The fitness function of the evolutionary search is evaluated by tions and variables varies as each flowsheet optimization model
performing a local non-linear optimization of the operational is created on the fly when the evolutionary algorithm generates
parameters for each scenario for which the solver IPOPT (Waechter a new set of design variables and units are included or excluded.
and Biegler, 2006) is used. This is done to reduce the number of inactive equations and to
improve to numerical conditioning of the problem. The mathemat-
6.3. Problem definition ical formulation of the optimization problem is shown in Eq. (6).
The evolutionary algorithm minimizes the expected value of the
The optimization variables and the uncertainties of the opti- specific production cost, approximated by the weighted sum over
mization problem of case study 1 are summarized in Table 2. A the ˝ scenarios. The investment cost is the same for all scenarios,
typical design candidate model has about 2000 equations and 5 and is thus not subscripted by the scenario index ω.
operational degrees of freedom (recourse variables). The recourse
1  COperation,ω + (CInvest /10)
˝
variables are: the temperature of the reactor TReactor , the partial
pressure of carbon monoxide in the reactor pCO , the partial pressure minimize CPT = (6)
˝ ṁProduct,Capacity,ω
of hydrogen in the reactor pH2 the molar flow of dodec-1-ene in the ω=1

Fig. 7. Sensitivities of the cost function regarding the uncertainties.

208 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217

Fig. 8. Operating windows of the recourse decisions.

subject to: The objective function is flat with many solutions having similar
cost values.
Process models (7)
The cost for the different scenarios of the 6 most promising
Purity constraint (8) designs that were identified by the optimization is shown in Fig. 5.
While the average values for these designs do not differ much,
Capacity constraint (9)
a decision is possible when the solutions for each scenario are
Three uncertainties were considered in this case study: the pre- examined. Designs 1 (both crystallizers exist) and 2 (no crystal-
exponential factor K0 that describes the equilibrium constant for lizers exist) are more expensive even in their best case scenario
the distribution of the catalyst between the two phases in the than the worst case scenarios of designs 3–6 (in these only the first
decanter, the price of steam and the price of cooling water. The crystallizer exists), so that it can be assumed that these designs are
upper and lower bounds for the uncertainties are given in Table 2. dominated and not part of the set of optimal designs.
The uncertain parameters are discretized in three levels, which As the best four solutions share this configuration, only one crys-
implies that 27 scenarios must evaluated for each design. As no tallizer before the product column will be considered. The best four
additional information was available, each scenario was assumed designs differ only slightly regarding the other design variables (the
to have the same probability of occurrence. The termination criteria number of stages and the location of the feeds of the distillation
for the evolutionary algorithm are: a maximum wall-clock time of columns). Therefore the design with the lowest average value was
48 h, a maximum number of 100 generations that have been evalu- chosen (design 6) which is shown in Fig. 6.
ated or that the objective has not been improved for 50 consecutive The first-order sensitivity of the cost function for design 6 with
generations. The time limit was reached after 37 generations. For respect to the uncertainties were determined with a multiple linear
this case study a total number of 382 designs were examined using regression. The results of this analysis are shown in Fig. 7. Although
the hybrid evolutionary algorithm in FSOpt (10 solutions for each the uncertainty in the catalyst loss parameter has an influence on
generation, 10 solutions for the initialization and 2 solutions of the the cost function, it is overshadowed by the uncertainty in the
partially evaluated 38th generation). This means that 10,314 (382 prices of cooling water and steam. Therefore predicting the dis-
designs × 27 scenarios) non-linear sub-problems were solved. tribution of the catalyst between the phases in the decanter more
precisely does not contribute much to the final decision and is not
6.4. Numerical results and discussion necessary at this point. Instead, the uncertainty in the cost is mostly
determined by the external influences of the prices for the utilities.
The progress of the evolutionary algorithm is shown in Fig. 4. The The parameters of design 6 are given in Table 3. It has to be noted
fitness (equivalent to the expected value of the stochastic problem) that both columns are rather small and the feed stage is located
improved over the course of the generations but a large stalling very high, especially in the solvent column. This suggests that for
period is apparent, where no better solution was found. As the this separation task a simple flash evaporation might suffice.
optimization was aborted after 48 h, the best solution is only an In Fig. 8 the maximum and minimum values of the most impor-
estimate of the true optimum, for which no guarantees can be made. tant recourse variables are shown (over all scenarios and scaled
 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217 209

with respect to the bounds). The partial pressures of the compo-

nents of the synthesis gas in the reactor are modified considerably
over the scenarios, while the temperatures in the reactor and the
decanter and the feed flow of nC12 are not changed much. This
interaction across the flowsheet can be explained by an examina-
tion of the solutions to the scenario subproblems. In those scenarios
in which the prices for cooling water and steam are high, the opti-
mal operation of the reactor leads to a reduced conversion (by about
1%) compared to the nominal case. This changes the composition
of the stream that leaves the reactor and the separation task is eas-
ier and the reflux and boilup ratios can be reduced. Therefore the
energy demand is lower in both columns, which compensates some
of the impact of the prices on the overall cost. Design 6 provides
enough operational freedom to almost eliminate the effect of the
uncertainties on the cost by varying the operating conditions.

7. Case study 2: purification of formic acid

7.1. Process description

Fig. 9. Superstructure of the purification process for formic acid. The second case study concerns a part of the Kemira–Leonard-
process for the production of formic acid from methyl-formate.
Table 3 Formic acid is a chemical used the textile, chemical and pharma-
Optimal design parameters for design 6. ceutical industry. The worldwide installed capacity was in the range
Process step Variable Unit Value of 330,000 t/a in 1988 (Reutemann and Kieczka, 2012). The overall
Kemira–Leonard-process is quite large, featuring several distilla-
Solvent column NStages [–] 17
Product column NStages [–] 17 tion columns and reaction steps. This case study is a first step in
Solvent column Feed stage [–] 16 developing an equation-based model for the optimization of entire
Product column Feed stage [–] 13 process in the FSOpt modelling framework. The goal of this case
Crystallizer 1 Exists [–] 1 study is to highlight the results of using stochastic programming,
Crystallizer 2 Exists [–] 0
Reactor Volume m3 30
and how they can be exploited in order to estimate the feasibil-
HX Area m2 20 ity of process alternatives in early stages of the process design
with parametric uncertainty. The main challenge arising from this

Fig. 10. Sensitivity screening of the uncertain parameters for case study 2.
210 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217

Table 4
Decisions of the optimization problem for case study 2.

Process step Variable Type Unit Min Max

Column 1 Stages above Design-disc [–] 5 15

Column 2 Stages above Design-disc [–] 5 15
Column 1 Stages below Design-disc [–] 5 15
Column2 Stages below Design-disc [–] 5 15
Column 1 Reboiler HX area Design-cont m2 5 600
Column 1 Condenser HX area Design-cont m2 5 300
Column 2 Reboiler HX area Design-cont m2 5 300
Column 2 Condenser HX area Design-cont m2 5 300
Column 1 Diameter Design-cont m 2.5 5
Column 2 Diameter Design-cont m 2.5 5
Wilson aH2 O,FA Uncertain [–] 1.102 1.347
Wilson aFA,H2 O Uncertain [–] 0.122 0.149
Heat transfer Coefficient Uncertain W/m2 /K 900 1100
Column 1 Reflux ratio Recourse [–] 0.01 10
Column 1 Boilup ratio Recourse [–] 0.01 10
Column 2 Reflux ratio Recourse [–] 0.01 10
Column 2 Boilup ratio Recourse [–] 0.01 10

case study is the implementation of the non-ideal thermodynam- 7.2. Model description
ics in the equation-based framework of FSOpt and the subsequent
solution of the optimization problem. The column model used here has a slightly different structure
The process part examined in this case study concerns the purifi- than the one in the first case study. Instead of optimizing the num-
cation of a mixture of formic acid (FA), water (H2 O), methyl-formate ber of stages and the feed stage separately, the column model was
(MeF) and methanol (MeOH) which usually takes place as the separated into a rectifying section and a stripping section. The num-
last step in the Kemira–Leonard-process. The process features two ber of trays in each section was optimized independently, removing
distillation columns operated at different pressures to achieve a the need for the safeguard that checks if the feed tray location is
separation despite the existence of an azeotrope between the two smaller than the maximum number of trays. With this formulation
components. The superstructure of this sub-process is shown in every combination of discrete design parameters encodes a feasible
Fig. 9. solution. Therefore the problem-specific repair algorithm was not

Fig. 11. Evolution of the fitness function over the generations for case study 2.
 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217 211

Fig. 12. Scenario production cost of the best 5 designs for case study 2.

needed any more. The column model was enhanced to be able to sieve trays and estimates the free-on-board cost including nozzles
describe the non-ideal behaviour of the liquid phase and to consider and access holes:
the azeotrope between water and formic acid.  1,5
0.53
In this case study, both columns were assigned a maximum hD
CInv,Column = 545, 000$ · (16)
number of trays of Nmax = 31 stages, including the feed stage (i.e. 100
15 trays above and below the feed). The numbers where obtained
in off-line studies using short-cut-models to identify the minimum The heat exchangers for the condensers and reboilers of the
number of stages, and were adjusted with an additional factor to columns were also included in the cost model. The investment costs
provide enough freedom in the design optimization. of the heat exchangers are also estimated using a correlation from
The non-ideal behaviour of the liquid phase that leads to the for- Woods (2007), which is reproduced in Eq. (17).
mation of an azeotrope between FA and H2 O is captured by using
 A 0.71
the Wilson equation (Eqs. (10)–(11)) for the calculation of the activ-
CInv,HX = 70, 000$ · (17)
ity coefficients. The non-ideality of the gas phase is neglected in this 100
case study.
  The total investment cost is calculated as the sum of the invest-
  ment of the columns and the heat exchangers.
1−ln xj i,j − 
xk k,i

x 
j j k,j
i = e j k
(10) CInvestmest = CInv,Column1 + CInv,Column2 + CInv,Reb,Column1
bi,j
i,j = eai,j + T (11) + CInv,Reb,Column2 + CInv,Con,Column1 + CInv,Con,Column2

In Eqs. (10) and (11)  i is the activity coefficient for a component, (18)
i, j are chemical component indices and k,j are the interaction
coefficients regressed on experimental data.
As this case study only concerns a small part of the entire
Kemira–Leonard-process and the reaction step is neglected, a sur-
rogate objective function was defined which tries to emulate the
most important cost factors. The goal of the optimization is to min-
imize the cost of the separation per tonne of product, as defined
in Eq. (12). In this equation, the cost per ton is defined by the sum
of the annual operation cost and the annualized investment cost,
divided by the annual capacity of the plant:

1  COperation,ω + (CInvest /10)

˝
minimize CPT = (12)
˝ ṁProduct,Capacity,ω
ω=1

subject to:

Process models (13)

Purity constraint (14)

Capacity constraint (15)

The investment cost of a distillation column is estimated using

a correlation from Woods (2007), which is reproduced in Eq. (16).
Here, h denotes the column height in m and D denotes the column
diameter, also in m. This correlation assumes a tray column with Fig. 13. Flow sheet of the best design for case study 2.
212 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217

The operation cost was estimated using the condenser and The model can be easily adapted for automated initializa-
reboiler duties of the columns. The equation for calculating the tion. For example, the column model includes a parameter IsIdeal,
operating cost is given in Eq. (19): which controls the actual implementation of the activity coefficient
model. If it is set to 1, then a dummy model is included which always
COperation,Column = PSteam · (QReboiler,Column1 + QReboiler,Column2 ) returns the value 1 for the activity coefficient, whereas the equa-
tions of the Wilson model are included when it is set to 0. Using this
+ PCoolingWater · (QCondenser,Column1
modelling technique, the model is first solved using ideal thermo-
+ QCondenser,Column2 ) (19) dynamics to generate consistent starting values for temperatures
and concentrations on the trays, and only afterwards the detailed
The MESH equations were implemented in FSOpt using Model- model is solved.
ica. The hierarchical model is transformed into a flat optimization The uncertain parameters where identified using a Monte-Carlo
model using FSOpt and solved using the IPOPT C# API. The solver screening of nine possible candidates. In this screening nine param-
is used with an automated initialization strategy based on empiri- eter candidates were randomly varied and the impact on the
cal observations. In this strategy IPOPT is first allowed to run for a objective function of the base case process (with 10 + 1 +10 stages
maximum of 300 iterations with a constant objective function and for each column) was measured. For each parameter the sensitiv-
relaxed constraints in order to find better starting values. After this ity was calculated using multiple linear regression. The parameter
initialization, the main optimization is performed with an iteration sensitivities are shown in Fig. 10. From these nine only the three
limit of 5000 and tighter constraint violation. All scenario prob- most sensitive were chosen to be included in the actual design opti-
lems are initialized with initial values determined by the solution mization. These are the Wilson interaction parameters aH2 O,FA and
of a surrogate optimization problem in which all uncertainties were aFA,H2 O and the heat transfer coefficient of the heat exchangers.
fixed at their expected values. The equations of the MESH column The three most sensitive uncertain parameters are discretized in
model can be found in Appendix A.4. three levels, which means that 27 scenarios are evaluated for each
7.3. Problem definition design. As no additional information was available, each scenario
was assumed to have an equal probability of occurrence. The three
The variables and the uncertainties of the optimization prob- termination criteria for the evolutionary algorithm are: a maxi-
lem of the second case study are summarized in Table 4. A typical mum wall-clock time of 48 h, a maximum number of 20 generations
design candidate model has about 2000 equations and 4 opera- have been evaluated or the objective has not been improved for 10
tional degrees of freedom. Again, the number of equations and consecutive generations. The optimization was finished after 20
variables varies as each flowsheet optimization model is created generations without reaching the time limit. For this case study
on the fly when the evolutionary algorithm generates a new set of a total number of 210 designs were examined using the hybrid
design variables and units are included or excluded. evolutionary algorithm in FSOpt (10 solutions for each genera-
A abbreviated form of the Modelica implementation of the opti- tion and 10 solutions for the initialization). Therefore 5670 (210
mization problem showing the flowsheet connectivity is given in designs × scenarios) non-linear sub-problems were solved.
listing 1. Omitted from this code fragment are the definition of the
objective function, the declaration of the degrees of freedom and 7.4. Numerical results and discussion
the cost model. The names of the units correspond to the names
used in the graphical superstructure given in Fig. 9. Using the The progress of the evolutionary algorithm is shown in Fig. 11.
object-oriented modelling paradigm, composing flowsheets from The fitness generally improved over the course of the generations
different process units is very straight forward due to the avail- but the optimization stalled between the generation 6 and 15,
ability of high-level modelling concepts like the connect-keyword, where no better solution was found. Again, during the last 6 gen-
which represents a stream connection between two units. erations no better solution could be found. Similar to the first case
Listing 1: The flowsheet model implemented in FSOpt-Modelica
 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217 213

Fig. 14. Operating windows of the best design for case study 2.

Fig. 15. Oversizing of the continuous design parameters for the best design for case study 2.
214 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217

Table 5 which the objective function was evaluated for optimized oper-
Optimal design parameters for design 6.
ational parameters for each scenario. The optimization of the
Process step Variable Unit Value operational parameters was performed by the non-linear program-
Column 1 NStages [–] 30 ming solver IPOPT.
Column 2 NStages [–] 26 The results of both case studies demonstrated that promising
Column 1 Feed stage [–] 16 designs can be identified with the help of two-stage stochastic pro-
Column 2 Feed stage [–] 11 gramming and that the impact of the uncertainties on the process
Column 1 Reboiler HX area m2 464.20
performance can be quantified. In both case studies, the uncertain-
Column 1 Condenser HX area m2 208.18
Column 2 Reboiler HX area m2 221.66 ties were modelled using a scenario approach with a three-level
Column 2 Condenser HX area m2 147.50 discretisation. In the first case study, the process for the hydro-
Column 1 Diameter m 2.89 formylation on dodec-1-ene was optimized and it was observed
Column 2 Diameter m 4.11
that the uncertainties in the separation part of the plant could be
largely compensated by varying the reactor operating conditions.
In the second case study, a process for the production of formic
study, the objective function here is very flat with many solutions acid was modelled in FSOpt and solved by the hybrid evolution-
having similar average values. ary algorithm. The distillation models were more complex, with
The set of the 5 most promising designs identified by the opti- non-ideal thermodynamics in the form of the Wilson equation for
mization is shown in Fig. 12. The production cost for each scenario calculating the liquid phase activities. It was observed here as well
is plotted over the different designs. that the uncertainties in the chemical properties can mostly be
The ability of IPOPT to solve the model for the second case study compensated by the recourse.
was very sensitive to the uncertain parameters. In some scenarios, The presented approach for handling the uncertainties by
which are not the same for every design, IPOPT converged to local decomposing the scenarios and solving them sequentially is suited
optima, which in some cases are far from the expected value (the for large problems with discrete design decisions when the appli-
spikes in Fig. 12). In these scenarios the reflux ratio assumed very cation of an evolutionary algorithm is beneficial. The first stage is
large values leading to very high energy costs. These outlier sce- optimized by a generate-and-test meta-heuristic, while the flow-
narios where omitted in the later analysis. The cost of the best 5 sheet mass and energy balances are solved with a state-of-the-art
solutions are very close to each other. Designs 3 and 4 show more non-linear solver. This approach combines the good features of both
variation in the cost depending on the scenario, whereas design 1 approaches. Due to the fact that the evolutionary search opera-
is very robust regarding the uncertainties (excluding the outliers tes derivative-free, the discrete design decisions can be optimized
caused by numerical instabilities). as integer variables, and the flowsheet can be pruned during the
In the following, only design 1 will be discussed further, as this search to reduce degeneracies and unused equations. The non-
design has the lowest cost in each scenario (disregarding the out- linear solver IPOPT proved to be capable of solving the resulting
liers) compared to any other design. The design parameters for medium-scale non-linear problems quite reliably.
design 1 are summarized in Table 5. The corresponding flowsheet The computation time for solving the design optimization prob-
is given in Fig. 13. lem is still quite large. Over 98% of the overall time is spent on
The operating windows of the recourse parameters for design solving the NLP subproblems. Under these conditions, optimization
1 are shown in Fig. 14. The most important recourse parameter algorithms requiring fewer evaluations of the designs might be bet-
is the reflux ratio in the second column which is varied most to ter suited. But several points favour the use of a meta-heuristic as
compensate the disturbances introduced by the scenarios. a first-stage solver.
As the continuous design parameters were optimized in the evo-
lutionary search, no guarantee can be made for optimality. In order 1. The ability to include non-relaxable integrality constraints
to estimate the quality of the solution for the continuous design greatly simplifies modelling. Instead of having to include relax-
parameters, the minimum required values where calculated in the ations for every tray in a column, the number of trays can directly
NLP model for all scenarios and the relative overdesign was evalu- be optimized without the need for binary existence variables
ated (see Fig. 15). For nearly all design parameters, the value chosen for each tray. Therefore the distillation column model is only
by the evolutionary search was minimal for at least one scenario. concerned with modelling the physical phenomena and not the
In these cases the relative overdesign is 1, which means that the superstructure connectivity. The topology is handled in the flow-
parameter was at the lowest possible value before incurring infea- sheet optimization problem in which the number of trays can be
sibility. The area of the condenser heat exchanger in column 1 and independently optimized. This separation of concerns increases
the area of the reboiler in column 2 could be further reduced with the reusability of the models in a different context.
minimal overdesigns of 1.88 and 1.43, then further reducing the 2. The generate-and-test structure of meta-heuristic algorithms
investment cost. generates a sequence of solution candidates and produces feasi-
ble solutions very early. The optimization can be aborted at any
8. Conclusions point in time and a feasible solution is (in most cases) available
already at the first iteration.
8.1. Discussion 3. Due to the strict stage decomposition, the NLP model is gener-
ated only before its evaluation, and all superfluous equations
This contribution presented the application of two-stage of non-existing units are removed. The reduced number of
stochastic programming to two medium-size flowsheets of chem- equations and variables reduces the computational burden and
ical processes, one with complex kinetics and distillation columns increases the robustness of the solver.
modelled by equilibrium stages and possibly additional crystalliz-
ers, the other featuring two distillation columns with non-ideal The main drawback of meta-heuristics is the inability to prove
liquid phases. Both case studies were solved with a stage decom- the true global optimum. But considering the fact, that the prob-
position approach implemented in a hybrid evolutionary algorithm lems dealt with in this framework are highly non-convex, local
which combined an evolutionary search for the optimization of dis- solvers cannot guarantee global optimality anyway. According to
crete and continuous design parameters (first stage variables) in the knowledge of the authors, global optimization algorithms are
 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217 215

still limited to more simplified (Ballerstein et al., 2011) or small- activation energy EA,r using a reference temperature for scaling
scale (Misener et al., 2014) problems. Sometimes special algorithms purposes.
can reduce the prohibitively large run times of standard solvers
k1,0 cnC12 CH2 cCO
(Karuppiah and Grossmann, 2008), but this is still a field of research. r1 = (A.1)
1 + K1,1 cnC12 + K1,2 cnC13 + K1,3 cH2
The full-factorial sampling of the uncertainty space leads to a
large number of scenarios if many parameters are considered, but


k2,0 cnC12 − (ciC12 /Kp,2 )
due to the decomposition, each problem is quite small. In the two r2 = (A.2)
1 + K2,1 cnC12 + K2,2 ciC12
case studies the algorithm was able to obtain good solutions which


were feasible for all scenarios. k3,0 cnC12 CH2 − (cC12an /Kp,3 )
Using the object-oriented language Modelica, flowsheet prob- r3 = (A.3)
1 + K3,1 cnC12 + K3,2 cC12an + K3,3 cH2
lems can be modelled concisely with a high degree of model-reuse.
The tool supports the automatic transformation of the object- r4 = k4,0 ciC12 cH2 (A.4)
oriented model into a flat, equation-based model for implemen-
tation in various mathematical programming solvers. r5 = k5,0 ciC12 cH2 cCO (A.5)
 1 
−EA,j 1
8.2. Outlook kj,0 = kj0 exp − , j = 1, . . ., 5 (A.6)
R T 378.15

As mentioned earlier, the second case study only concerned a

A.2. Decanter model
small sub-process of the entire Kemira–Leonard-process for the
production of formic acid. The overall goal is to implement the other
The decanter model is represented as a equilibrium-stage model
parts of this process, most importantly the reaction step, and to ana-
which calculates the output molar fraction using experimentally
lyze the impact of the uncertainties on the units and on the recycle
determined equilibrium distribution coefficients. With correlations
structure of the larger process. Due to the need for further distilla-
to calculate the equilibrium constants Kc for a given temperature
tion columns, this model will become much larger in size, making
T and p (the operating point of the unit) the equilibrium-based
it challenging to solve the flowsheet model for multiple scenarios.
separation unit model consists of the following equations.
The second topic for future research is the handling of large
processes under uncertainty. In order to achieve an acceptable yc ṅliquid2 = zc ṅfeed − xc ṅliquid1 (A.7)
approximation of the uncertainties, a large number of scenarios
has to be used to reduce the impact of sampling errors. While the yc = Kc (T, p)xc (A.8)
expected value of a stochastic problem can be estimated with a The equilibrium coefficients are based on experimental equi-
small number of samples, the extreme scenarios are difficult to find, librium data taken from Brunsch (2013), to which an exponential
especially when the uncertainties enter into of non-linear func- function (Eq. (A.9)) was fitted.
tions. Methods for efficient sampling have already been discussed  K1

by Diwekar and Kalagnanam (1997), Helton and Davis (2003) and K = exp K0 + + K2 · T (A.9)
Helton et al. (2005). Future work will concern the investigation of T
methods for sampling and the identification of critical scenarios
and a reduction of the number of the scenarios that must be eval- A.3. Crystallizer model
uated in order to get an accurate estimation of the distribution of
the cost. The crystallizer model used in the hydroformylation case study
is only valid for the modelling of melt crystallization processes. In
melt crystallization a mixture is cooled below the melting point
Acknowledgements
of the desired component, which starts to precipitate. Sometimes
another component with a similar melting points also precipitates
The financial support of Deutsche Forschungsgemeinschaft for
and the molecules of the impurity are included in the crystal struc-
the Sonderforschungsbereich Transregio 63 (SFB/TR 63) is grate-
ture.
fully acknowledged. The authors would like to thank all co-workers
The following equations govern the behaviour of the compo-
in InPROMPT for their contributions to the knowledge base that
nent split. The model is based on the short-cut model presented by
led to the definition of the superstructure of the hydroformylation
Franke et al. (2008). In this model it is assumed that the crystallizing
process.
component forms an eutectic mixture with the impurity and can be
obtained as a (nearly) pure crystal. The location of the eutectic point
Appendix A. Process models is modelled by a simple exponential correlation. The melting point
Tmelt,Crystal , the enthalpy of melting Hmelt and the integral impurity
In the following sections, the models for the main process units distribution coefficient Kint are taken from (Beierling and Ruether,
used in the case studies are given. 2012). The energy required for obtaining the crystal is estimated
using the correlation of Wellinghoff and Wintermantel (1991). In
A.1. Reactor model of the hydroformylation process Eq. (A.14), QAPP /Qm and QPUMP /Qm are the specific energy duties for
cooling the apparatus and for pumping the liquid. The values for
A reduced chemical reaction path system considering the hydro- these parameters are also taken from the cited article.
formylation (r1 and r5 ), the isomerization (r2 ) and the hydration
yc ṅSolid = zc ṅFeed − xc ṅLiquid (A.10)
(r3 and r4 ) of the olefins is considered. The kinetics used in this
case study that define the rates r1 to r5 are taken from (Hentschel xSolid,Impurity = Kint · xFeed,Impurity (A.11)
et al., 2014) and are reproduced in Eqs. (A.1)–(A.6). The molar con-
centrations ci are given in mol/L. Several constants Kr,i are used xSolid,Crystal = 1 − xSolid,Impurity (A.12)
in the inhibition terms, which are defined in the cited article. The   
Hmelt 1 1
pre-exponential factors kr,0 for each reaction r are calculated with xLiquid,Crystal = exp − (A.13)
R Tmelt,Crystal T
an Arrhenius equation from the reaction rate constant kr0 and the
216 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217

 QAPP QPUMP

Summation and equilibrium
QCryst = ṅSolid · Hmelt · 1+ + (A.14)
Qm Qm

Nc

Nc
A.4. Distillation model yi,j − xi,j = 0, i = 0, . . ., N + 1 (A.30)
j j
The MESH equations describe a distillation column as a cascade
of interconnected equilibrium stages and form one large non-linear yi,j − i (T, xi )Ki,j (T, p, xi )xi,j = 0, j ∈ Nc , i = 0, . . ., N + 1 (A.31)
equation system which has to be solved simultaneously. Many for-
mulations of the MESH equations exist. In this contribution the
formulation from Biegler (2010) is used. The MESH equations are References
reproduced in Eqs. (A.15)–(A.31). The capital letters B and D denote
the molar flows in the bottom and the distillate. L and V denote Acevedo J, Pistikopoulos EN. A parametric MINLP algorithm for process synthesis
problems under uncertainty. Ind Eng Chem Res 1996];35:147–58.
the liquid and vapour molar flow on each stage. The molar frac-
Baharev A, Neumaier A. Chemical process modeling in Modelica. In: Proceedings of
tions are given by x for the liquid phase and y for the vapour phase. the 9th International Modelica Conference; 2012]. p. 955–62.
The energy transferred on each stage is denoted by the variable Ballerstein M, Kienle A, Kunde C, Michaels D, Weismantel R. Towards global opti-
Q. The enthalpies HL and HV are calculated using linear correlations mization of combined distillation–crystallization processes for the separation
of closely boiling mixtures. In: 21st European Symposium on Computer Aided
regressed to literature data. The equilibrium constants Ki,j are calcu- Process Engineering, vol. 29; 2011].
lated via the Antoine vapour pressure correlation and the constant Behr A, Henze G, Schomaecker R. Thermoregulated liquid/liquid catalyst separation
pressure in the column (neglecting pressure drop). The total num- and recycling. Adv Synth Catal 2006];348(12–13):1485–95.
Beierling T, Ruether F. Separation of the isomeric long-chain aldehydes dodecanal/2-
ber of stages is denoted by N and the feed stage enters the column methylundecanal via layer melt crystallization. Chem Eng Sci 2012];77:71–7.
at LocFeed . In the first case study the activity coefficient  is assumed Biegler LT. Nonlinear programming. SIAM; 2010].
to be 1 for all components. In the second case study it is calculated Birge JR, Louveaux F. Introduction to stochastic programme. Springer Series in Oper-
ations Research; 1997].
by the Wilson equation. Brunsch Y. Dissertation Temperaturgesteuertes Katalysatorrecycling für die
Total mass balances homogen katalysierte Hydroformylierung langkettiger Alkene Dissertation. TU
Dortmund; 2013].
B + V0 − L1 = 0 (A.15) Caroe CC, Schultz R. Dual decomposition in stochastic Integer programming. Oper
Res Lett 1998];24:37–45.
Li + Vi − Li+1 − Vi−1 = 0, i ∈ [1, N], i  Loc Feed (A.16) Diwekar UM, Kalagnanam JR. Efficient sampling technique for optimization under
uncertainty. AIChE J 1997];43(2):440–7.
Dowling AW, Balwani C, Gao Q, Biegler LT. Equation-oriented optimization of
Li + Vi − Li+1 − Vi−1 − F = 0, i ∈ Loc Feed (A.17)
cryogenic systems for coal oxycombustion power generation. Energy Proc
2014];63:421–30.
LN+1 + D − VN = 0 (A.18) Dowling AW, Biegler LT. A framework for efficient large scale equation-oriented
flowsheet optimization. Comput Chem Eng 2015];72:3–20.
Degrees of freedom Eiben AE, Smith JE. Introduction to evolutionary computing (natural computing
series). Springer; 2008].
LN+1 Franke MB, Nowotny N, Ndocko EN, Gorak A, Strube J. Design and optimization of a
RReflux = (A.19)
D hybrid distillation/melt crystallization process. AIChE J 2008];54:2925–42.
Friedman Z, Ingalls J, Siirola JD, Watson JP. Block-oriented modeling of superstruc-
V0 ture optimization problems. Comput Chem Eng 2013];57:10–23.
RBoilup = (A.20) Grossmann IE, Caballero JA, Yeomans H. Mathematical programming approaches to
B
the synthesis of chemical process systems. Korean J Chem Eng 1999];16:407–26.
Component mass balances Grossmann IE, Caballero JA, Yeomans H. Advances in mathematical programming
for the synthesis of process systems. Latin Am Appl Res 2000];30:263–84.
Bx0,j + V0 y0,j − L1 x1,j = 0, j ∈ Nc (A.21) Grossmann IE, Sargent RWH. Optimum design of chemical plants with uncertain
parameters. AIChE J 1978];24:1021–8.
Hart WE, Watson JP, Woodruff D. Pyomo: modeling and solving mathematical pro-
Li xi,j + Vi yi,j − Li+1 xi+1,j − Vi−1 yi−1,j = 0, grams in Python. Math Prog Comput 2011];3(3):219–60.
(A.22) Helton JC, Davis FJ. Latin hypercube sampling and the propagation of uncertainty in
j ∈ Nc , i ∈ [1, N], i  Loc Feed analyses of complex systems. Reliab Eng Syst Safety 2003];81(1):23–69.
Helton JC, Davis FJ, Johnson JD. A comparison of uncertainty and sensitivity analysis
Li xi,j + Vi yi,j − Li+1 xi+1,j − Vi−1 yi−1,j − FxF,j = 0, results obtained with random and Latin hypercube sampling. Reliab Eng Syst
(A.23) Safety 2005];89(3):305–30.
j ∈ Nc , i ∈ Loc Feed Hentschel B, Peschel A, Freund H, Sundmacher K. Simultaneous design of the
optimal reaction and process concept for multiphase systems. Chem Eng Sci
(LN+1 + D)xN+1,j − VN yN,j = 0, j ∈ Nc (A.24) 2014];115:69–87.
Holmgren M. Process simulation in industrial projects. In: Proceedings of the 3rd
xN+1,j − yN,j = 0, j ∈ Nc (A.25) international Modelica conference; 2003]. p. 353–8.
Karuppiah R, Grossmann IE. Global optimization of multiscenario mixed integer
nonlinear programming models arising in the synthesis of integrated water
Enthalpy balances networks under uncertainty. Comput Chem Eng 2008];32:145–60.
Khor CS, Chachuat B, Shah N. A superstructure optimization approach for water net-
BH B + V0 HV,0 − L1 HL,1 − QR = 0 (A.26) work synthesis with membrane separation-based regenerators. Comput Chem
Eng 2012];42:48–63.
Kondili E, Pantelides CC, Sargent RWH. A general algorithm for short-term
Li HL,i + Vi HV,i − Li+1 HL,i+1 − Vi−1 HV,i−1 = 0, i ∈ [1, N], i  Loc Feed scheduling of batch operations. I. MILP formulation. Comput Chem Eng
1993];17(2):211–27.
(A.27) Lee S, Logsdon JS, Foral MJ, Grossmann IE. Superstructure optimization of the olefin
separation process. Comput Aided Chem Eng 2003];14:191–6.
Mauss J. Modelica instance creation. In: Proceedings of the 4th international Mod-
elica conference; 2005].
Li HL,i + Vi HV,i − Li+1 HL,i+1 − Vi−1 HV,i−1 − FH F = 0, i ∈ Loc Feed Misener R, Fuentes Garí M, Rende M, Velliou E, Panoskaltsis N, Pistikopoulos EN, et al.
Global superstructure optimisation of red blood cell production in a parallelised
(A.28) hollow fibre bioreactor. Comput Chem Eng 2014];71:532–53.
Olsson H, Otter M. Modelica – a unified object-oriented language for systems
modeling – language specification – version 3.3. Linköping, Sweden: Modelica
Association; 2012].
Pistikopoulos EN, Ierapetritou MG. Novel approach for optimal process design under
VN HV,N − (LN+1 + D)HL,D − QC = 0 (A.29) uncertainty. Comput Chem Eng 1995];19(10):1089–110.
 J. Steimel, S. Engell / Computers and Chemical Engineering 81 (2015) 200–217 217

Quaglia A, Gargalo CL, Chairakwongsa S, Sin G, Gani R. Systematic network synthesis editors. 16th European symposium on computer aided process engineering and
and design: problem formulation, superstructure generation, data management 9th international symposium on process systems engineering; Elsevier; vol. 21
and solution. Comput Chem Eng 2014]. of computer aided chemical engineering; 2006]. p. 1967–72.
Recker S, Skiborowski M, Redepenning C, Marquardt W. A unifying framework for Till J, Sand G, Engell S, Emmerich M, Schönemann L. A hybrid algorithm for
optimization-based design of integrated reaction–separation processes. Comput solving two-stage stochastic integer problems by combining evolutionary algo-
Chem Eng 2015]. rithms and mathematical programming methods. Comput-Aided Chem Eng
Reutemann W, Kieczka H. Formic acid. In: Ullmann’s encyclopedia of industrial 2005];20A:187–92.
chemistry. Wiley-VCH; 2012]. Till J, Sand G, Urselmann M, Engell S. A hybrid evolutionary algorithm for solving
Rooney WC, Rooney WC, Biegler LT, Biegler LT. Optimal process design with model two-stage stochastic integer programs in chemical batch scheduling. Comput
parameter uncertainty and process variability. AIChE J 2003];49(2). Chem Eng 2007];31(5-6):630–47.
Sahinidis N. Optimization under uncertainty: state-of-the-art and opportunities. Tometzki T, Engell S. Risk conscious solution of planning problems under uncer-
Comput Chem Eng 2004];28:971–83. tainty by hybrid multi-objective evolutionary algorithms. Comput Chem Eng
Skiborowski M, Harwardt A, Marquardt W. Efficient optimization-based design 2010];35:2521–39.
for the separation of heterogeneous azeotropic mixtures. Comput Chem Eng Urselmann M, Barkmann S, Sand G, Engell S. A memetic algorithm for global opti-
2014];72:34–51. mization in chemical process synthesis problems. IEEE Trans Evol Comput –
Steimel J, Harrmann M, Schembecker G, Engell S. Model-based conceptual design Special Issue Adv Memetic Comput 2011];15:659–83.
and optimization tool support for the early stage development of chemical pro- Waechter A, Biegler LT. On the implementation of a primal-dual interior point filter
cesses under uncertainty. Comput Chem Eng 2013];59:63–73. line search algorithm for large-scale nonlinear programming. Math Program
Steimel J, Harrmann M, Schembecker G, Engell S. A framework for the modeling 2006];106(1):25–7.
and optimization of process superstructures under uncertainty. Chem Eng Sci Wellinghoff G, Wintermantel K. Schmelzkristallisation – theoretische Vorausset-
2014];115:225–37. zungen und technische Grenzen. Chem Ing Tech 1991];63(9):881–91.
Tanvir M, Mujtaba I. Optimisation of design and operation of MSF desalination pro- Woods DR. Rules of thumb in engineering practice. Wiley-VCH; 2007].
cess using MINLP technique in gPROMS. Desalination 2008];222(1–3):419–30. Yeomans H, Grossmann IE. A systematic modeling framework of super-
Till J, Sand G, Engell S. Stochastic integer programming in chemical batch sched- structure optimization in process synthesis. Comput Chem Eng 1999];23:
uling: evolution strategies vs. exact algorithms. In: Marquardt W, Pantelides C, 709–31.